|Publication number||US7899271 B1|
|Application number||US 11/778,391|
|Publication date||Mar 1, 2011|
|Filing date||Jul 16, 2007|
|Priority date||Sep 15, 2004|
|Also published as||EP2171642A2, EP2171642A4, WO2009048660A2, WO2009048660A3|
|Publication number||11778391, 778391, US 7899271 B1, US 7899271B1, US-B1-7899271, US7899271 B1, US7899271B1|
|Inventors||Darin S. Williams|
|Original Assignee||Raytheon Company|
|Export Citation||BiBTeX, EndNote, RefMan|
|Patent Citations (15), Referenced by (7), Classifications (26), Legal Events (2)|
|External Links: USPTO, USPTO Assignment, Espacenet|
This application is a continuation-in-part of co-pending U.S. application Ser. No. 10/941,203 entitled “FLIR-to-Missile Boresight Correlation and Non-Uniformity Compensation of the Missile Seeker” and filed on Sep. 15, 2004 now U.S. Pat. No. 7,463,753 and claims priority under 35 U.S.C. 120.
This invention was made with Government support under Contract HQ0006-01-C-0001/101616 awarded by the Ballistic Missile Defense Organization. The Government has certain rights in this invention.
1. Field of the Invention
This invention relates to the determination of non-uniformity compensation (NUC) terms for optical imagers, and more particularly to a low-cost method of calibrating precision NUC terms based on a moving target.
2. Description of the Related Art
Optoelectronic imagers such as Focal Plane Arrays (FPAs) in the IR, near visible, visible or other bands detect incident radiation and convert it to electrical signals to record an image of a scene. The response of the imager on a pixel-by-pixel basis can change dramatically and non-uniformly over time and based on environmental and operating conditions. These non-uniformities appear as fixed-pattern noise in the recorded images. The purpose of non-uniformity correction; known as ‘calibration’ if done off-line typically during manufacture or as ‘compensation’ if done on-line just prior to use of the imager, is to reduce the fixed-pattern noise.
Although arbitrary order non-uniformity compensation (NUC) terms can be computed to correct for non-uniformities, the terms of most interest are typically the offset (0th order term) and the gain (1st order term). The offset terms are relatively unstable, hence are typically compensated in the field just prior to or as the imager is being used. Gain terms are relatively stable and thus are typically calibrated offline, usually at the time of manufacture. In some systems, the gain terms may be ‘tweaked’ just prior to use. Known techniques for compensating both the offset and gain terms are based on the premise that on-average all pixels should see the same value.
The predominant approach for compensating the offset terms uses a blurred version of the scene created optically, through motion of the imager, or through temporal averaging in the field. Based on this premise, any high spatial frequency components that are detected in the blurred image for each pixel are deemed to be the result of non-uniform pixel response. The blurred image is corrected to remove the high frequency components. The same correction is then applied to the subsequent non-blurred image. This approach is serviceable for relatively “flat” imagery but struggles with scenes which contain significant content at high spatial frequencies. These may be perceived as non-uniformities and “compensated” producing scene and body-motion dependent artifacts.
The predominant approach for calibrating the gain terms is to expose the imager to uniform flood sources at different temperatures. The response of each pixel is taken as the difference value between these two sources to first cancel pixel offsets. A gain term is then calculated for each pixel to flatten the apparent response over the entire imager. The offset terms are discarded but the gain terms are saved to calibrate the imager. Different sets of gain terms may be calibrated for different operating conditions of the imager. A problem with this approach is that the flood measurements are conducted in a separate vacuum test chamber from the moving target tests performed with the calibrated imager. Providing multiple vacuum test chambers and moving the imager between test chambers significantly increases test time and cost.
The present invention provides a system and method for precision calibration of non-uniformity compensation (NUC) terms including but not limited to the 1st order gain terms. The described approach allows for the use of the same test chamber to perform non-uniformity calibration and to perform the moving target tests.
The current approach works by scanning a test scene having a target and a background at different illumination intensity levels in an overlapping pattern across the imager field of view (FOV) and cross-referencing multiple measurements of each pixel of a test scene as viewed by different pixels in the imager; each imager pixel (to be fully compensated) sees multiple different scene pixels and each scene pixel is seen by multiple imager pixels. This approach is based on the simple yet novel premise that every pixel in the array that looks at the same thing should see the same thing.
In an embodiment, a method of non-uniformity calibration (NUC) for a pixilated imager comprises providing a test scene having a target in a background at different intensity levels. Light from the scene is collimated and scanned across the imager FOV to generate a sequence of target images in which the test scene is repeated in an overlapping pattern at different target placements within the imager FOV so that each imager pixel to be fully compensated images at least one target intensity level and at least one background intensity level (and preferably multiples of the nominal target and background levels respectively). The sequence of target images is registered to the target and filtered (e.g. median or weighted average filter of the intensity levels) to form an estimate of the test scene (registered average image). This estimate represents an estimate of the ideal response assuming perfect NUC terms. Since each point on the target is seen by multiple imager pixels, variations in the pixel response tend to cancel out. This cancellation is imperfect but improves with iteration of the process. This estimate is backward mapped to the different target placements in the imager FOV to create an estimated target image for each target image. At this point, each imager pixel to be fully compensated will have associated with it data pairs (target estimate, target measured) for at least one and preferably multiple background scene pixels and at least one and preferably multiple target scene pixels. For each imager pixel, an Nth order correlation is performed on the data pairs for different scene pixels to estimate non-uniformity calibration (NUC) terms (typically gain and offset) for each imager pixel such that all fully-compensated imager pixels provide substantially the same response to the same input intensity level. The NUC terms can be applied to the target images and the process repeated to refine the calculation of the NUC terms. Typically, the gain terms will be saved as calibration weights for the imager and the offsets discarded.
In another embodiment, in order to fully compensate pixels near the edge of the imager FOV, the test scene is scanned across and beyond the imager FOV so that the edge pixels image at least one and preferably multiple background and target intensity levels. The sequence of target images will now include ‘interior’ target images in which the target lies fully within the imager FOV and ‘edge’ target images in which only a portion of the target (less than the whole target) lies within the imager FOV. These edge target images tend to be more difficult to register accurately and can induce tracking error. Therefore, in one embodiment only the interior target images are used to create the registered average image during the 1st and subsequent iterations. Once the NUC terms for the pixels that can be fully-compensated using only the interior target images have converged then the edge target images are registered to the last scene estimate and used to generate NUC terms for all of the fully compensated pixels. This process can be enhanced by keeping track of the variability in the forward mapping (creating the estimate of the test scene) and/or backward mapping (estimating the ideal target images based on that scene estimate) and weighting those fully-compensated imager and/or scene pixels with less variability more heavily. This has the effect of favoring those imager pixels with the most consistent response for estimating scene content during the “forward mapping” and of favoring those scene pixels which are most spatially and temporally consistent in estimating pixel responses during “backward mapping”. More specifically, imager pixels are weighted to form the registered average image and scene pixels are weighted when performing the Nth order correlation of the data to compute the NUC terms and possibly when performing the variance calculation to determine the imager pixel variability.
In an embodiment, to create the estimate of the test scene during at least the 1st iteration each scene pixel is subjected to a non-linear filter such as a median filter that rejects at least some outlier intensity levels. These outlier scene pixels often correspond to bad, dead or blinker imager pixels. Removing these imager pixels up front has been found to speed convergence of the NUC terms. After the 1st iteration, an imager pixel stationarity map may be computed based on differences for each imager pixel between the measured and estimated pixel response for different scene pixels and used to assign weights to each imager pixel. In general the weights are inversely proportional to the variance of an imager pixel response. Dead or blinker pixels will typically demonstrate a large variance and be marginalized. The weights are then used to weight the contributions of the imager pixels during the formation of the registered average image (i.e., the scene estimate).
In another embodiment, the estimate of the test scene is spatially masked based on a priori knowledge of the test scene to create the estimated target images. The mask excludes areas of high spatial variability such as transitions between the target and background intensity levels and includes areas of low variability such as the interior of the target and an annular region in the background around the target. The mask effectively sets the weights in the excluded areas to zero and the weights in the included areas to one. Alternately, the mask can be used to explicitly set the weights in the excluded areas to zero and not be directly applied to the image data. For additional refinement, a gradient of the pixel values can be computed for the included areas. If individual pixel gradient values are too high (e.g. exceeds a threshold) they are also excluded. The remaining scene pixels in the included areas can be left with uniform weights or their weights can be assigned by computing the weighted variance at each scene pixel in the estimate of the test scene, i.e. the weighted variance of the registered imager pixels that contribute to each scene pixel, to form a scene pixel stationarity map. These scene pixel weights are then used to weight the contributions of the scene pixels when the Nth order correlation is performed to fit the estimated and measured pixel response data for each imager pixel. The weights may also be used to weight the contributions of the difference values when computing the variance for the imager pixel stationarity map.
These and other features and advantages of the invention will be apparent to those skilled in the art from the following detailed description of preferred embodiments, taken together with the accompanying drawings, in which:
The present invention describes a system and method for precision calibration of non-uniformity compensation (NUC) terms including but not limited to the 1st order gain terms. The described approach allows for the use of the same test chamber to perform non-uniformity calibration and to perform the moving target tests. The current approach works by scanning a test scene having a target and a background at different illuminations in an overlapping pattern across the imager FOV and cross-referencing multiple intensity level measurements of each pixel of a test scene as viewed by different pixels in the imager; each imager pixel to be fully compensated (hereinafter referred to as a “fully-compensated imager pixel”) sees multiple different scene pixels and each scene pixel is seen by multiple fully compensated imager pixels. This approach is based on the simple yet novel premise that every pixel in the array that looks at the same thing should see the same thing. Depending upon the application not every imager pixel need be fully compensated e.g. pixels near the edges of the imager, and thus the method of calibration applies to the ‘fully-compensated imager pixels’ where fully compensated means at least gain compensation and not merely offset compensation. Although designed for calibration at the time of manufacturing, the current approach could be adapted to perform calibration or compensation in the field based on a known moving target, e.g. by scanning the bright moon against black space.
As shown in
A scanning mirror 24 scans the test scene across the field of view (FOV) 23 of an imager 26 to generate a sequence of target images 30 in which the test scene is repeated in an overlapping pattern 28 at different target placements across the imager FOV so that each fully-compensated imager pixel 32 images at least one target intensity level and at least one background intensity level (step 34). The scan pattern or FOV motion 38 across the scene is tracked (step 40). The scan pattern is suitably a standard back-and-forth raster scan, typically sawtooth, or boustraphedonic for somewhat greater efficiency, but can be any pattern that provides the requisite overlap. The preferred amount of scene overlap is actually considerably greater than that depicted in
In a currently preferred embodiment, in order to fully compensate pixels 37 near the edge of the imager FOV 23, the test scene is scanned across and beyond the imager FOV, typically at least 50% of one test scene beyond each edge, so that all pixels image at least one and preferably multiple background and target intensity levels. The sequence of target images will now include ‘interior’ target images 36 in which the target lies fully within the imager FOV and ‘edge’ target images 39 in which only a portion of the target lies within the imager FOV. These edge target images tend to be more difficult to register accurately and can induce tracking error. Therefore, in one embodiment only the interior target images 36 are used to create the registered average image during the 1st and subsequent iterations. Once the NUC terms for the pixels that can be fully-compensated using only the interior target images have converged then the edge target images 39 are registered to the last scene estimate and used to provide additional pair data to generate NUC terms for all of the pixels.
The sequence of target images 30 is fed to a computer 42 that is configured to compensate each target image (step 44) with a priori NUC terms 46 if any are provided. A priori NUC terms could by typical average offset and gain terms found for similar imagers. The computer processes the sequence of compensated target images to generate a scene estimate, extract estimates of the ideal target images from the scene estimate and fit the measured response data to the corresponding estimates for each imager pixel to provide the NUC terms for the imager (step 48). The computer may output the NUC terms 50, specifically the gain terms, for calibration (step 52) after a single iteration or the computer may apply the NUC terms to the original target images (step 44) and repeat the process one or more times to recalculate the NUC terms. As described above, in the case that the imager FOV is over scanned in order to fully compensate all pixels, the computer may identify and only process ‘interior’ target images until the last iteration at which time the edge target images are incorporated into the NUC terms that are calculated for all fully compensated imager pixels.
By starting successive iterations with better compensated target images, the computer process provides a better scene estimate, hence better estimates of the ideal target images. As a result, the fit of the estimated response to the measured response of the original target images will produce better NUC terms for each fully-compensated imager pixel. Typically, 5 or 6 iterations are sufficient for the NUC terms to converge to the point of vanishing returns. In principle a smart cutoff could be done based on the degree of change in successive iterations; in practice it has not proven necessary. Note, the processing algorithm could be configured to fit the estimate to the compensated target image to produce residual NUC terms that would be accumulated to provide the full NUC terms. This process is typically sufficient for 0th order terms but the gain and higher order terms are more sensitive to error accumulation. Once the NUC terms are provided to calibrate the imager, moving target testing on the imager can be performed in the same test chamber (step 53).
The basic process implemented by computer 42 to generate the scene estimate, estimated target images and estimated NUC terms for a single iteration is illustrated in
Computer 42 receives the compensated target images 60 generated in step 44 and registers them to the target 12 in the test scene (step 62). In general, a sufficient number of properly placed images must be registered to provide desired imager pixels that satisfy the criterion to be fully-compensated. Typically, all of the interior target images will be registered. As mentioned previously, in the case of over scanning the imager FOV to fully compensate pixels near the edges of the FOV, the edge target frames may be incorporated and registered at this point but better results are achieved by only incorporating them to provide additional data pairs to calculate the final NUC terms. The computer then filters the registered images 60 to generate an estimate of the ideal test scene as if the imager pixel NUC terms were perfectly calibrated (referred to herein as the “registered average image” 64) (step 66). The region of registered average image 64 outside the area of the test scene 10 (i.e. the target and a limited portion of the background) is suitably windowed and discarded for further processing. At this point, the computer suitably normalizes the registered average image (step 68) to prevent the offset terms from drifting if this process is iterated. The image is suitably normalized by forcing the average background pixel intensity level to zero and then forcing the average difference between target pixel intensity level and the background pixel intensity level to a constant. This serves to fix the contrast of the images.
The registered average image 64 is backward mapped to the different target placements in the imager FOV to create an estimated target image 69 for each target image (step 70). Backward mapping amounts to using the FOV motion to copy the registered average image 64 back to the appropriate location in the imager so that the target image 30 and the estimated target image 69 are properly registered. If the regions outside the test scene are not discarded during processing, backward mapping amounts to cutting out the appropriate full-sized image from the collage (registered average image). Each estimated target image is an estimate of what the imager pixel responses should have been assuming perfect calibration.
At this point, each fully-compensated imager pixel will have associated with it data pairs 72 (target estimate, target measured) that represent the estimated ideal and measured pixel response for different scene pixels including at least one and preferably multiple background scene pixels, and at least one and preferably multiple target scene pixels, e.g. the required samples to satisfy the criterion for fully-compensated. The multiple target pixels suitably have intensity levels that lie within a first range about a nominal target level and the multiple background pixels suitably have intensity levels that lie within a second range about a nominal background level. For each fully-compensated imager pixel, computer 22 performs an Nth order correlation (step 74) on the data pairs to estimate non-uniformity calibration (NUC) terms 50 for the imager to make the measured data fit the estimated data such that all fully-compensated pixels provide substantially the same response to the same input intensity level. In general, all lower order terms are provided by an Nth order correlation. For an arbitrary Nth order correlation, the target will require N different nominal intensity levels; hence a 1st order correlation to extract gain terms only needs a single target intensity level. A 1st order correlation generates both the offset and gain terms. For example, to a 1st order E=M*Gain+Offset (or equivalently E=(M−Offset)*Gain) where M is the measured pixel response, E is the estimate of the ideal pixel response and the Gain is the slope of the linear approximation. The 1st order correlation entails, for example, performing a standard minimum mean squared error (MMSE) fit of the data to this equation to estimate the Gain and Offset terms 76 and 78, respectively.
The enhanced process which entails keeping track of the variability in the forward mapping and weighting those fully compensated imager pixels with less response variability more heavily to create the estimate of the test scene in the form of registered average image 64 is illustrated in
To further improve convergence and the precision of the NUC terms, during at least the 1st iteration (before the imager pixel stationarity map can be updated) instead of applying a uniform set of weights the computer preferably applies a non-linear filter such as a median filter 86 to each scene pixel to reject at least some outlier values. These outlier scene pixels often correspond to bad or dead imager pixels. Removing these values up front has been found to speed convergence of the NUC terms. The median filter may be used for more than just the 1st iteration until the stationarity map has converged to identify the good and bad imager pixels.
The enhanced process which entails keeping track of the variability in the backward mapping and assigning larger weights to those scene pixels with less variability is illustrated in
To further improve convergence and the precision of the NUC terms, a priori knowledge of the test scene, i.e. areas that are likely to exhibit low and high variability and particular intensity levels, may be used to mask the scene pixels to remove areas that are likely to exhibit high variability and to select the number of target and background pixels (step 93). As shown in
The mask can be applied during step 93 either to the scene pixel stationarity map 92 or to the registered average image 64 (or to the image data at any point thereafter). In the former case, all of the image data is processed as before and the weights are applied when performing the Nth order correlation or when computing the imager pixel variance. The weights in the excluded regions are zero so the scene pixel data is ignored. The weights in the included regions may be forced to one (or any constant) or allowed to vary as the process iterates based on individual scene pixel response described above. In the later case, the image data in the excluded regions is simply removed or ignored. This has two primary benefits. First, the computer does not have to continue to process scene pixel data from the excluded regions, which improves computational efficiency. Second, in step 68 only data from included regions is normalized. If the entire registered average image is normalized and then the excluded image data removed the normalization is likely to be off.
Following the later approach, masking excludes regions of the test scene so that the estimated target images 69 effectively only include pixel data on the interior of the target and in the annular ring around the target outside the transition area. The pixel data in the excluded transition area and remainder of the background is effectively gone. Consequently, unexplained difference images 79 will only include difference values in the interior and the annular ring. The excluded values could be carried forward but will not be used to compute imager pixel variance in step 82. Similarly, only pixel data in the included regions will be correlated to the measured pixel response data from the target images in step 74.
The overall effect of the forward and backward enhanced procedures is that iterating the process will identify ‘good’ and ‘bad’ imager pixels and rebalance the imager pixel weights accordingly and will identify ‘stationary’ and ‘non-stationary; portions of the scene, and will rebalance the scene pixel weights accordingly. Together (or separately) these enhanced procedures both speed and ensure convergence to precision NUC terms.
While several illustrative embodiments of the invention have been shown and described, numerous variations and alternate embodiments will occur to those skilled in the art. Such variations and alternate embodiments are contemplated, and can be made without departing from the spirit and scope of the invention as defined in the appended claims.
|Cited Patent||Filing date||Publication date||Applicant||Title|
|US5129595||Jul 3, 1991||Jul 14, 1992||Alliant Techsystems Inc.||Focal plane array seeker for projectiles|
|US5185526||Oct 31, 1990||Feb 9, 1993||Grumman Aerospace Corporation||Optical scene generator for testing infrared detector modules|
|US5420421||Oct 12, 1994||May 30, 1995||Hughes Aircraft Company||Wide dynamic range non-uniformity compensation for infrared focal plane arrays|
|US5903659||Apr 17, 1997||May 11, 1999||Raytheon Company||Adaptive non-uniformity compensation algorithm|
|US5940140||Sep 10, 1996||Aug 17, 1999||Ultimatte Corporation||Backing luminance non-uniformity compensation in real-time compositing systems|
|US5978162 *||Mar 19, 1997||Nov 2, 1999||Cirrus Logic, Inc.||Synchronous read channel integrated circuit employing a channel quality circuit for calibration|
|US6021011 *||Mar 19, 1997||Feb 1, 2000||Cirrus Logic, Inc.||Synchronous read channel employing a digital center frequency setting for a variable frequency oscillator in discrete time timing recovery|
|US6175407 *||Dec 17, 1998||Jan 16, 2001||Identix Incorporated||Apparatus and method for optically imaging features on the surface of a hand|
|US6452686 *||Apr 2, 2002||Sep 17, 2002||General Scanning, Inc.||Method and system for high speed measuring of microscopic targets|
|US6750974 *||Sep 16, 2002||Jun 15, 2004||Gsi Lumonics Corporation||Method and system for 3D imaging of target regions|
|US7035475||Jun 17, 1999||Apr 25, 2006||Raytheon Company||Non-traditional adaptive non-uniformity compensation (ADNUC) system employing adaptive feedforward shunting and operating methods therefor|
|US20010042825||Feb 16, 2001||Nov 22, 2001||Young Ching-Ju J.||Ditherless non-uniformity compensation for infrared detector arrays with recursive spatial low pass filtering|
|US20050213845||Mar 24, 2004||Sep 29, 2005||General Electric Company||Method and product for processing digital images|
|US20050242268||Dec 2, 2004||Nov 3, 2005||Oakley Philip J||Detection apparatus|
|US20060269155||May 9, 2006||Nov 30, 2006||Lockheed Martin Corporation||Continuous extended range image processing|
|Citing Patent||Filing date||Publication date||Applicant||Title|
|US8665334||Aug 22, 2011||Mar 4, 2014||Raytheon Company||Blur-calibration system for electro-optical sensors and method using a moving multi-target constellation|
|US8681880 *||Nov 5, 2012||Mar 25, 2014||Apple Inc.||Adaptive dithering during image processing|
|US9167239||Sep 12, 2013||Oct 20, 2015||Raytheon Company||System and moving modulated target with unmodulated position references for characterization of imaging sensors|
|US20110125441 *||May 26, 2011||Williams Darin S||Optical distortion calibration for electro-optical sensors|
|US20120114229 *||Jan 21, 2011||May 10, 2012||Guoqing Zhou||Orthorectification and mosaic of video flow|
|US20130064445 *||Nov 5, 2012||Mar 14, 2013||Apple Inc.||Adaptive Dithering During Image Processing|
|WO2013028758A1 *||Aug 22, 2012||Feb 28, 2013||Raytheon Company||Blur-calibration system for electro-optical sensors and method using a moving multi-target constellation|
|U.S. Classification||382/284, 382/275, 382/282, 382/274|
|International Classification||G06K9/32, H04N5/365, H04N5/367|
|Cooperative Classification||F41G7/2293, F41G7/002, H04N17/002, H04N5/3675, H04N5/3655, H04N5/33, G06T2207/30212, G06T2207/30208, G06T2207/20216, G06T2200/32, G06T5/50, G06T2207/10016, G06T5/008, G01S3/7803, F41G7/2253|
|European Classification||H04N5/33, H04N5/365A2, H04N17/00C, H04N5/367A|
|Jul 16, 2007||AS||Assignment|
Owner name: RAYTHEON COMPANY, MASSACHUSETTS
Free format text: ASSIGNMENT OF ASSIGNORS INTEREST;ASSIGNOR:WILLIAMS, DARIN S.;REEL/FRAME:019561/0628
Effective date: 20070713
|Aug 6, 2014||FPAY||Fee payment|
Year of fee payment: 4